Emit share_amount on lending events (closes #577)

[shubhamessier]

Append share_amount: WrappedI80F48 to LendingAccountDepositEvent,
LendingAccountWithdrawEvent, LendingAccountBorrowEvent, and
LendingAccountRepayEvent. Indexers can now reconstruct share-based
balances from event history without re-deriving cumulative share price
or polling on-chain state.

Implementation:

• New field is appended to each event struct (no reordering), preserving
  serialization compat for existing consumers per maintainer guidance in
  the issue thread.
• Share delta is captured at each emit site by snapshotting
  bank_account.balance.{asset,liability}_shares before and after the
  deposit / withdraw / borrow / repay call. Because each public wrapper
  uses a single-direction Balance{Increase,Decrease}Type, the orthogonal
  share field is invariant across the call, so the snapshot is equivalent
  to the delta computed internally — no *_balance_internal signature
  change required.
• Type is WrappedI80F48 (not u64 or f64) per maintainer comment:
  shares are fractional, f64 cast is CU-expensive, raw I80F48 preserves
  full precision.
• Sign convention: positive value = shares moved (minted for
  deposit/borrow, burned for withdraw/repay). *_all paths produce the
  full pre-balance amount since post-close balance is zero.

Touched all event construction sites (12 total): the 4 in
instructions/marginfi_account/, plus juplend / drift / kamino / solend
deposit and withdraw, since they construct the same event types.

Verification:

• cargo check -p marginfi --no-default-features --features custom-heap
  passes clean.
• Anchor build / SBF build / basic-tests not run in dev environment
  (toolchain unavailable); CI should validate

[shubhamessier]

Hello @jgur-psyops, can you kindly have a look at these changes once, thanks!

implementation per your guidance in #577:

Decisions you flagged:

1. Append-only ordering: confirmed, all 4 events get the new field at the
   end. No existing field reordered.
2. I80F48 vs f64: went with WrappedI80F48. Preserves full precision and
   avoids the CU hit you mentioned on the f64 cast. Happy to switch to
   f64 if downstream ergonomics outweigh that single-line change in
   events.rs plus .to_num() at the 12 emit sites.